ArXiV ML/AI/CV papers summary

We are currently witnessing a profound shift in artificial intelligence. We are moving away from the era of “black-box” models—those that merely predict the next token or classify pixels—and into an era of agentic, grounded, and physically-aware systems. This transition marks a move from “model-centric” development, obsessed with parameter counts, to “system-centric” engineering, where the focus is on reliability, accountability, and the integration of AI into the complex, real-world workflows of science and industry.

Theme 1: Grounding Intelligence in Physical and Scientific Reality

The frontier of AI is moving into the physical world, where models must respect the laws of physics, chemistry, and biology. We are no longer satisfied with models that “hallucinate” plausible-sounding answers; we now demand that they respect the fundamental constraints of our universe.

• Physical Latent Structuring: Researchers are tackling “physical representation laziness,” where models predict outcomes without understanding the underlying mechanics. Spectral-Target Physical Latent Structuring for JEPA-Style World Models uses Fourier auxiliary heads to force latent spaces to adhere to physical properties. Similarly, TourPhysics: Bringing Physics to World Models for Exploration and Manipulation from a Single Image and AccidentSim: Generating Vehicle Collision Videos with Physically Realistic Collision Trajectories from Real-World Accident Reports integrate physical simulators into the generative loop to ensure outputs are physically consistent.
• Scientific Discovery: AI is becoming a “scientist-in-the-loop.” Hakken: Predicting future discoveries to fill the gaps in today’s knowledge uses knowledge graphs to predict undocumented scientific relationships. In the physical sciences, Fast Surrogate Modeling of Excitable and Oscillatory FitzHugh-Nagumo Dynamics with Parametric Neural Operators and An Energy-Based Conservative-Dissipative Latent Neural Evolution Operator for Magnetization Dynamics replace expensive numerical solvers with neural operators. Furthermore, Mapping Dark-Matter Clusters via Physics-Guided Diffusion Models and Towards AI-Driven Nanomedicine Discovery: A Benchmark and Multimodal Learning Framework for Nano Self-Assembly Prediction demonstrate how AI can turn costly trial-and-error processes into predictive computational tasks.
• Multimodal Synergy: Synergistic Information Disentanglement for Omni-modal Slide Representation Learning in Computational Pathology uses information theory to force models to learn the irreducible synergy between disparate data types, such as histology and genomics, preventing “modality collapse.”

Theme 2: The Agentic Turn and Structural Accountability

We are transitioning from static chatbots to autonomous agents capable of executing code and navigating complex environments. This shift necessitates a move toward “structural accountability,” where we verify the reasoning process rather than just the final output.

• Verification and Grounding: To combat hallucinations, researchers are deconstructing LLM outputs into verifiable claims. GRACE: Graph-Grounded Reflective Agent Copilot Engine for Expert-in-the-Loop Knowledge Expansion and Enoki: Efficient Multi-Level Hallucination Detection ground agent outputs in knowledge graphs. For tool use, Persistent Teacher Anchoring for Tool-Using Agents and SiLR: Structure-Preserving Admission and Process Reward for LLM Tool Agents introduce “gates” to prevent unsafe tool execution.
• Neuro-Symbolic Reasoning: Think-Verify-Revise: Neuro-Symbolic Visual Reasoning with Vision-Language Models and Dynamic Logic Tensor Networks creates a feedback loop where models propose logical rules that are verified by differentiable logic networks. This procedural approach is mirrored in Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action Manipulation and From Intent to Evidence: Policy-Steered Multi-Strategy Retrieval for Long-Video Agents, which emphasize the need for an “evidence ledger” to maintain state in long-horizon tasks.
• Evaluation: The field is maturing its standards with frameworks like Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation, $\tau^\tau$-Bench: An Environment for End-To-End, Realistic Agent Construction, and Measuring AI Accountability Through Argumentation Analysis: Can Model Reasoning Withstand Scrutiny?, which treats the quality of an agent’s defense as a proxy for its reliability.

Theme 3: Efficiency, Compression, and Inference-Time Scaling

As models grow, the “reasoning economy” becomes paramount. We are seeing a shift toward “algorithm-system co-design,” where models are optimized for the physical constraints of the hardware they inhabit.

• Extreme Compression: Deep Microcompression: Structured Pruning and Bit-packed Quantization for Microcontrollers achieves 55.8x compression for microcontrollers. Similarly, DenseScout: Algorithm-System Co-design for Budgeted Tiny Object Selection on Edge Platforms and Evolving Layer-Specific Scalar Functions for Hardware-Aware Transformer Adaptation optimize architectures for edge accelerators.
• Inference Optimization: BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference and KVMem: Virtualizing Million-Token Agent Workspaces on a Consumer GPU manage memory bottlenecks. SPD: Single Pass Decoding for Generative Reranking and ACE: Adaptive Calibration-Free Expert Skipping for MoE-based LLMs accelerate inference by reducing redundant computation.
• Efficient Training: Scale-QLoRA: Code-Invariant Adapter Merging for Native 4-bit Microscaling LLMs and RISE: Recursive Improvement via Self-Extrapolating Policy Distillation demonstrate that high-level reasoning can be achieved by distilling “hard” examples rather than relying on massive datasets.

Theme 4: Interpretability, Robustness, and Sociotechnical Governance

Deploying AI in the real world requires more than just accuracy; it requires transparency, fairness, and the ability to adapt to unpredictable environments.

• Mechanistic Interpretability: Locating and Steering Refusal Beyond Attention and The Struggle Between Continuation and Refusal map the internal “geography” of models to understand behaviors like refusal. ProToMEx: Rapid, Interpretable Explanations via Structured Representations and SMILE: Self-Explainable Multimodal Information Bottleneck for Medical Diagnosis force models to produce explanations intrinsically tied to their decisions.
• Robustness and Adaptation: Cross-dataset transportability of pediatric chest X-ray deep learning across three countries: discrimination, calibration, operating-point failure, and limited-label recovery highlights the dangers of relying on internal accuracy alone. To mitigate this, Test-Time Adaptation via Cache Personalization for Facial Expression Recognition in Videos and FailureSpot: Label-Efficient Timestamp-Level Failure Detection for Vision-Language-Action Models enable models to detect their own failures and adapt in real-time.
• Sociotechnical Governance: As AI enters human-centric domains, we must address bias and trust. A Fairness Audit of the Duckworth-Lewis-Stern Method, Counterfactual Fairness Audits of Multi-Step Clinical LLM Agents, and Labels have Human Values: Value Calibration of Subjective Tasks emphasize that we must calibrate systems to respect the pluralistic nature of human values, ensuring that our intelligent systems remain equitable partners.